Cardiac Segmentation And Myocardial Ischemia Auxiliary Diagnosis Based On Deep Learning

In recent years,with the advancement of the aging population and the induction of poor living and eating habits,the morbidity and mortality of cardiovascular disease(CVD)in China has started to increase year by year,and the prevention,control and treatment of CVD is urgently needed.Medical image segmentation plays an important role in the early screening and diagnosis of CVD,and the powerful feature representation capability of deep learning makes it more and more closely integrated with medical images,and medical image segmentation and assisted diagnosis based on deep learning plays an increasingly important role in the clinical treatment of CVD.This paper mainly deals with the design and optimization of semi-supervised segmentation algorithms based on cardiac MRI,as well as myocardial segmentation and visualization-aided diagnosis of myocardial ischemia based on perfusion CT.In terms of semi-supervised segmentation algorithm optimization,the development of deep learning high-performance segmentation models often requires supervised training based on a large number of labeled data sets.How to effectively utilize the large amount of unlabeled raw data stored in data centers is an urgent problem to be solved in the implementation of deep learning algorithms.As for myocardial ischemia assisted diagnosis,perfusion CT,as a noninvasive imaging diagnostic tool,can be used for accurate ischemia diagnostic analysis through the calculation of blood flow reserve fraction.The accurate segmentation of myocardial segments provides a powerful tool for visual assessment and quantitative analysis of specific regions of the myocardium.In addition,the proposed framework for effective training of semisupervised segmentation maximizes the potential of myocardial segmentation models trained on limited labeled data.The contribution of the work in this paper consists of the following three main parts:(1)A semi-supervised cardiac segmentation training framework,Adversarial Semisupervised Segmentation Framework based on U-shape Network Combined with Multiple Attention and Convolution(SSMAC-UNet),is proposed to address the problem that large-scale annotation of medical images is difficult to acquire.Framework based on U-shape Network Combined with Multiple Attention and Convolution,SSMAC-UNet).Firstly,a Transformer global feature-enhanced segmentation network model MAC-UNet is designed,and an adversarial semi-supervised training framework based on a refined full convolution discriminator network is further proposed using the proposed network as a baseline network.The expressive potential of the proposed framework in the heart segmentation task is effectively demonstrated,and the average Dice coefficient of the proposed semi-supervised training framework is 0.9035 at a supervised data share of 0.5,which exceeds most of the segmentation network models based on fully supervised training.(2)A cardiac segmentation and bull’s-eye map visualization framework that follows international myocardial segmentation standards is proposed for the problem of aiding diagnosis of myocardial ischemia based on perfusion CT.Based on the 3D U-Net myocardial segmentation model and anatomical key point detection model,this framework can quickly complete the automatic segmentation of 17 myocardial segments and map the myocardial blood flow intensity distribution to a 2D bull’s-eye diagram based on the segmentation results,which can assist physicians to quickly determine the possible ischemic regions and the lesions of the corresponding blood supplying coronary arteries and make further diagnostic plans.Based on the consistent statistical analysis of two gold standards of myocardial ischemia,invasive flow reserve fraction and digital subtraction angiography,the high statistical correlation between the diagnostic results based on the proposed framework and the gold standard was verified.The framework is expected to provide an intelligent and reliable diagnostic aid for the work related to myocardial ischemia analysis.(3)To address the problem that the traditional medical image visualization interaction system can hardly meet the cross-platform applications,a web-based heart segmentation and visualization interaction system is designed and implemented based on the above-mentioned research work on supervised and semi-supervised heart segmentation algorithms.The system supports cardiac image segmentation based on multiple segmentation algorithms,visualization and rendering of segmentation results,as well as the localization of suspicious areas and calculation of indicators.It is convenient for doctors to analyze and diagnose patient image data anytime and anywhere,give corresponding treatment plan,and reduce the risk of misdiagnosis of cardiovascular diseases.This paper investigates the assisted diagnosis and segmentation algorithms for cardiovascular diseases based on cardiac magnetic resonance imaging and perfusion CT.A segmentation network model and a semi-supervised training framework are proposed to address the problem of difficulty in obtaining large-scale annotations for medical images.Additionally,a myocardial segment segmentation and bull’s eye map visualization framework are proposed to assist doctors in diagnosing myocardial ischemia.Finally,a web-based heart segmentation and visualization interactive system is implemented to facilitate doctors in analyzing and diagnosing patient image data and providing corresponding treatment plans.